Main conference: abstracts
CO-Fun: A German Dataset on Company Outsourcing in Fund Prospectuses for Named Entity Recognition and Relation Extraction
Neda Foroutan, Markus Schröder, Andreas Dengel
The process of cyber mapping gives insights in relationships among financial entities and service providers. Centered around the outsourcing practices of companies within fund prospectuses in Germany, we introduce a dataset specifically designed for named entity recognition and relation extraction tasks. The labeling process on 948 sentences was carried out by three experts which yields to 5,969 annotations for four entity types (Outsourcing, Company, Location and Software) and 4,102 relation annotations (Outsourcing–Company, Company–Location). State-of-the-art deep learning models were trained to recognize entities and extract relations showing first promising results. An anonymized version of the dataset, along with guidelines and the code used for model training, are publicly available at https://doi.org/10.5281/zenodo.12745116.
Lex2Sent: A bagging approach to unsupervised sentiment analysis
Kai-Robin Lange, Jonas Rieger, Carsten Jentsch
Unsupervised text classification, with its most common form being sentiment analysis, used to be performed by counting words in a text that were stored in a lexicon, which assigns each word to one class or as a neutral word. In recent years, these lexicon-based methods fell out of favor and were replaced by computationally demanding fine-tuning techniques for encoder-only models such as BERT and zero-shot classification using decoder-only models such as GPT-4. In this paper, we propose an alternative approach: Lex2Sent, which provides improvement over classic lexicon methods but does not require any GPU or external hardware. To classify texts, we train embedding models to determine the distances between document embeddings and the embeddings of the parts of a suitable lexicon. We employ resampling, which results in a bagging effect, boosting the performance of the classification. We show that our model outperforms lexica and provides a basis for a high performing few-shot fine-tuning approach in the task of binary sentiment analysis.
Discourse-Level Features in Spoken and Written Communication
Hannah J. Seemann, Sara Shahmohammadi, Manfred Stede, Tatjana Scheffler
Using PARADISE, a German corpus of thematically parallel blog posts and podcast transcripts, we test how reliably a document’s original mode can be classified based on discourse-level features only. Our results show that classifying mode with a document’s distribution of discourse relations as well as the frequency of discourse connectives and discourse particles is possible and informative of the nature of these document types. We provide our dataset annotated with discourse relations (Rhetorical Structure Theory), German discourse connectives, and discourse particles.
Semiautomatic Data Generation for Academic Named Entity Recognition in German Text Corpora
Pia Schwarz
An NER model is trained to recognize three types of entities in academic contexts: person, organization, and research area. Training data is generated semiautomatically from newspaper articles with the help of word lists for the individual entity types, an off-the-shelf NE recognizer, and an LLM. Experiments fine-tuning a BERT model with different strategies of post-processing the automatically generated data result in several NER models achieving overall F1 scores of up to 92.45%.
GERestaurant: A German Dataset of Annotated Restaurant Reviews for Aspect-Based Sentiment Analysis
Nils-Constantin Hellwig, Jakob Fehle, Markus Bink, Christian Wolff
We present GERestaurant, a novel dataset consisting of 3,078 German language restaurant reviews manually annotated for Aspect-Based Sentiment Analysis (ABSA). All reviews were collected from Tripadvisor, covering a diverse selection of restaurants, including regional and international cuisine with various culinary styles. The annotations encompass both implicit and explicit aspects, including all aspect terms, their corresponding aspect categories, and the sentiments expressed towards them. Furthermore, we provide baseline scores for the four ABSA tasks Aspect Category Detection, Aspect Category Sentiment Analysis, End-to-End ABSA and Target Aspect Sentiment Detection as a reference point for future advances. The dataset fills a gap in German language resources and facilitates exploration of ABSA in the restaurant domain.
Revisiting the Phenomenon of Syntactic Complexity Convergence on German Dialogue Data
Yu Wang, Hendrik Buschmeier
We revisit the phenomenon of syntactic complexity convergence in conversational interaction, originally found for English dialogue, which has theoretical implication for dialogical concepts such as mutual understanding. We use a modified metric to quantify syntactic complexity based on dependency parsing. The
results show that syntactic complexity convergence can be statistically confirmed in one of three selected German datasets that were analysed.
Given that the dataset which shows such convergence is much larger than the other two selected datasets, the empirical results indicate a certain degree of linguistic generality of syntactic complexity convergence in conversational interaction.
We also found a different type of syntactic complexity convergence in one of the datasets while further investigation is still necessary.
OMoS-QA: A Dataset for Cross-Lingual Extractive Question Answering in a German Migration Context
Steffen Kleinle, Jakob Prange, Annemarie Friedric
When immigrating to a new country, it is easy to feel overwhelmed by the need to obtain information on financial support, housing, schooling, language courses, and other issues.
If relocation is rushed or even forced, the necessity for high-quality answers to such questions is all the more urgent.
Official immigration counselors are usually overbooked, and online systems could guide newcomers to the requested information or a suitable counseling service.
To this end, we present OMoS-QA, a dataset of German and English questions paired with relevant trustworthy documents and manually annotated answers, specifically tailored to this scenario.
Questions are automatically generated with an open-source large language model (LLM) and answer sentences are selected by crowd workers with high agreement.
With our data, we conduct a comparison of 5 pretrained LLMs on the task of extractive question answering (QA) in German and English. Across all models and both languages, we find high precision and low-to-mid recall in selecting answer sentences, which is a favorable trade-off to avoid misleading users. This performance even holds up when the question language does not match the document language. When it comes to identifying unanswerable questions given a context, there are larger differences between the two languages.
Few-Shot Prompting for Subject Indexing of German Medical Book Titles
Lisa Kluge, Maximilian Kähler
With the rise of large language models (LLMs), many tasks of natural language processing (NLP) have reached unprecedented performance levels.
One task where LLMs have not yet been evaluated on is subject indexing with a large controlled target vocabulary.
In this work, an LLM is applied to the task of subject indexing a dataset of German medical book titles.
The results are compared to two common baseline methods already in productive use at our institution.
One critical parameter in a few-shot prompting approach is the composition of examples given to the LLM for instruction.
To select examples, we apply two similarity measures between book title and gold-standard labels. We hypothesise that these notions of similarity can serve as a measure of difficulty of the task.
We find that the LLM does not outperform the baselines.
Still, (off-the-shelf) LLMs can be a valuable addition in an ensemble of methods for subject indexing as they do not depend on training data.
Querying Repetitions in Spoken Language Corpora
Elena Frick, Henrike Helmer, Dolores Lemmenmeier-Batinić
In this paper, we present a tool for searching repetitions in interaction corpora. Our approach based on the MTAS-technology uses common search token indices to retrieve repetitions from spoken language transcripts in a dynamic way. The CQP Query Language and a graphical user interface menu with extensive settings specially designed for conversation analysis researchers allow to find repetitions of complex linguistic forms in various pragmatic contexts. Furthermore, the web application enables searching for repetition constructions that may contain synonyms and hyp(er)onyms coming from GermaNet or from custom-defined word lists uploaded to the tool.
A comparison of data filtering techniques for English-Polish LLM-based machine translation in the biomedical domain
Jorge del Pozo Lérida, Kamil Kojs, Janos Mate, Mikołaj Antoni Barański, Christian Hardmeier
Large Language Models (LLMs) have become state-of-the-art in Machine Translation (MT), often trained on massive bilingual parallel corpora scraped from the web, that contain low-quality entries and redundant information, leading to significant computational challenges. Various data filtering methods exist to reduce dataset sizes, but their effectiveness largely varies based on specific language pairs and domains. This paper evaluates the impact of commonly used data filtering techniques—LASER, MUSE, and LaBSE—on English-Polish translation within the biomedical domain. By filtering the UFAL Medical Corpus, we created varying dataset sizes to fine-tune the mBART50 model, which was then evaluated using the SacreBLEU metric on the Khresmoi dataset, having the quality of translations assessed by bilingual speakers. Our results show that both LASER and MUSE can significantly reduce dataset sizes while maintaining or even enhancing performance. We recommend the use of LASER, as it consistently outperforms the other methods and provides the most fluent and natural-sounding translations.
Linguistic and extralinguistic factors in automatic speech recognition of German atypical speech
Eugenia Rykova, Mathias Walther
Automatic speech recognition (ASR) has been already used in speech and language therapy, including diagnostic tasks and practice exercises for people with aphasia (PWA). The lack of relevant data makes it difficult to evaluate the algorithms’ suitability for German-speaking PWA. For the current project, four open-source ASR models were selected based on their performance on other types of atypical speech, and the details of their evaluation are presented in this paper. The four selected models are generally robust to speakers’ gender and age. The one-word recognition functions better for words of moderate length. Speech rate should be neither too slow nor too quick for lower error rates both in words and phrases, and the latter should be also of moderate length.
An Improved Method for Class-specific Keyword Extraction: A Case Study in the German Business Registry
Stephen Meisenbacher, Tim Schopf, Weixin Yan, Patrick Holl, Florian Matthes
The task of $\textit{keyword extraction}$ is often an important initial step in unsupervised information extraction, forming the basis for tasks such as topic modeling or document classification. While recent methods have proven to be quite effective in the extraction of keywords, the identification of $\textit{class-specific}$ keywords, or only those pertaining to a predefined class, remains challenging. In this work, we propose an improved method for class-specific keyword extraction, which builds upon the popular $\texttt{KeyBERT}$ library to identify only keywords related to a class described by $\textit{seed keywords}$. We test this method using a dataset of German business registry entries, where the goal is to classify each business according to an economic sector. Our results reveal that our method greatly improves upon previous approaches, setting a new standard for $\textit{class-specific}$ keyword extraction.
A Multilingual Dataset of Adversarial Attacks to Automatic Content Scoring Systems
Ronja Laarmann-Quante, Christopher Chandler, Noemi Incirkus, Vitaliia Ruban, Alona Solopov, Luca Steen
Automatic content scoring systems have been shown to be vulnerable to adversarial attacks, i.e. to answers that human raters would clearly recognize as incorrect or even nonsense but that are nevertheless rated as correct by an automatic system. The existing literature on this topic has so far focused on English datasets. In this paper, we present a multilingual dataset of adversarial answers for English, German, French and Spanish based on the multilingual ASAP content scoring dataset introduced by Horbach et al. (2023). We apply different methods of generating adversarial answers proposed in the literature, e.g. sampling n-grams from existing answers or generic corpora or inserting adjectives and adverbs into incorrect answers. In a baseline experiment, we show that the rate at which adversarials are rejected by a model depends on the adversarial method used, interacting with the language and the prompt-specific dataset a model was trained on.
Version Control for Speech Corpora
Vlad Dumitru, Matthias Boehm, Martin Hagmüller, Barbara Schuppler
While the audio recordings of a corpus represent the ground truth, transcriptions are -- in the case of manual annotations -- subject to human error, and subject to changes related to technology improvements underpinning automated annotation methods.
In order to facilitate the dynamic extension of speech corpora, we introduce a novel tool for centralized version control for speech corpora, enabling the automatic check-in and merging of annotations. It considers typical workflows of phoneticians, linguists and speech technologists, and enables the development of dynamic, collaborative, and perpetually-improving speech corpora.
Tabular JSON: A Proposal for a Pragmatic Linguistic Data Format
Adam Roussel
Existing linguistic data formats tend to be very general and powerful yet difficult to use on a day-to-day basis, so that practitioners often reach for underpowered ad-hoc text formats that require error-prone string parsing. We propose a pragmatic JSON-based linguistic data format that is flexible enough to cover most types of linguistic annotations and scenarios. It avoids the need for string parsing, as the serialized data representation is trivially convertible to tabular data structures that are immediately usable in data analysis applications.
Redundancy Aware Multiple Reference Based Gainwise Evaluation of Extractive Summarization
Mousumi Akter, Shubhra Kanti Karmaker Sant
The ROUGE metric is commonly used to evaluate extractive summarization task, but it has been criticized for its lack of semantic awareness and its ignorance about the ranking quality of the extractive summarizer. Previous research has introduced a gain-based automated metric called Sem-nCG that addresses these issues, as it is both rank and semantic aware. However, it does not consider the amount of redundancy present in a model summary and currently does not support evaluation with multiple reference summaries. It is essential to have a model summary that balances importance and diversity, but finding a metric that captures both of these aspects is challenging. In this paper, we propose a redundancy-aware Sem-nCG metric and demonstrate how the revised Sem-nCG metric can be used to evaluate model summaries against multiple references as well which was missing in previous research. Experimental results demonstrate that the revised Sem-nCG metric has a stronger correlation with human judgments compared to the previous Sem-nCG metric and traditional ROUGE and BERTScore metric for both single and multiple reference scenarios.
Complexity of German Texts Written by Primary School Children
Jammila Laâguidi, Dana Neumann, Ronja Laarmann-Quante, Stefanie Dipper, Mihail Chifligarov
While the development of children's literacy is of large interest to researchers, few studies have yet been based on corpora of children's texts. We investigate the development of text complexity in freely-written texts of German primary school children between 2nd and 4th grade based on the longitudinal Litkey Corpus (Laarmann-Quante et al., 2019b) using NLP methods. These texts are retellings of given picture stories. Although the picture stories may constrain the vocabulary and grammar, our hypothesis is that complexity increases over time. We measure complexity using various lexical and syntactic features. The results show that our hypotheses are largely confirmed but that there are outliers that might arise because some picture stories could be more stimulating than others.
Using GermaNet for the Generation of Crossword Puzzles
Claus Zinn, Marie Hinrichs, Erhard Hinrichs
Wordnets are playing an important role in research, but so far they have found little use in practical applications that are aimed at the general public. In this paper, we present a crossword generator that exploits lexical-semantic resources such as GermaNet. The software is capable of (i) automatically filling in the grid of a crossword puzzle with words taken from GermaNet for variable grid sizes, and (ii) generating clues for each word that is included in the grid. Crossword generation is not trivial, and we report on the effectiveness of various heuristic search functions that we have used.
A Crosslingual Approach to Dependency Parsing for Middle High German
Cora Haiber
This work presents the development and evaluation of a dependency parser for Middle High German Universal Dependencies utilising modern German as a support language for low-resource MHG. A neural dependency parser is trained with Stanza achieving UAS = 92.95 and LAS = 88.06. To ensure the parser’s utility in facilitating and speeding up manual annotation to build a scaling UD treebank of MHG, a thorough error analysis shows the model’s structural reliability as well as frequently confused labels. Hence, this work constitutes an effort to counterbalance the under-representation of historical languages in dependency treebanks and attend to the need of historical treebanks in contemporary linguistic research by utilising the UD extensions and accordingly annotated corpora published by Anonymous (2024).
Discourse Parsing for German with new RST Corpora
Sara Shahmohammadi, Manfred Stede
For RST-style discourse parsing in German, so far there has been only one corpus available and used, the single-genre Potsdam Commentary Corpus (PCC). Very recently, two new RST corpora of other genres have been made available. In our work, we build a homogeneously-annotated German RST corpus by changing the PCC annotations so they become compatible with the new corpora. We then run parsing experiments on different constellations of train/test splits over the three genres involved and report the results. A modified and streamlined version of the DPLP parser is prepared and made available, so that overall, the "resource situation" for German discourse parsing is notably improved.
Fine-grained quotation detection and attribution in German news articles
Fynn Petersen-Frey, Chris Biemann
The task of quotation detection and attribution deals with identifying quotation spans together with their associated role spans such as the speaker.
We describe an approach to solve the task of fine-grained quotation detection and attribution using a sequence-to-sequence transformer model with constrained decoding.
Our model improves vastly upon the existing baselines on the German news articles quotation dataset, thereby making it feasible for a first time to automatically extract attributed quotations from German news articles.
We provide an extensive description of our method, discuss alternative approaches, performed experiments using multiple foundation language models and method variants, and analyzed our model's prediction errors.
Our source code and trained models are available.
AustroTox: A Dataset for Target-Based Austrian German Offensive Language Detection
Pia Pachinger, Janis Goldzycher, Anna Maria Planitzer, Wojciech Kusa, Allan Hanbury, Julia Neidhardt
Model interpretability in toxicity detection greatly profits from token-level annotations. However, currently, such annotations are only available in English. We introduce a dataset annotated for offensive language detection sourced from a news forum, notable for its incorporation of the Austrian German dialect, comprising 4,562 user comments. In addition to binary offensiveness classification, we identify spans within each comment constituting vulgar language or representing targets of offensive statements. We evaluate fine-tuned Transformer models as well as large language models in a zero- and few-shot fashion. The results indicate that while fine-tuned models excel in detecting linguistic peculiarities such as vulgar dialect, large language models demonstrate superior performance in detecting offensiveness in AustroTox. We will soon publish the data and code.
OneLove beyond the field - A few-shot pipeline for topic and sentiment analysis during the FIFA World Cup in Qatar
Christoph Rauchegger, Sonja Mei Wang, Pieter Delobelle
The FIFA World Cup in Qatar was discussed extensively in the news and on social media. Due to news reports with allegations of human rights violations, there were calls to boycott it. Wearing a OneLove armband was part of a planned protest activity. Controversy around the armband arose when FIFA threatened to sanction captains who wear it. To understand what topics Twitter users Tweeted about and what the opinion of German Twitter users was towards the OneLove armband, we performed an analysis of German Tweets published during the World Cup using in-context learning with LLMs. We validated the labels on human annotations. We found that Twitter users initially discussed the armband's impact, LGBT rights, and politics; after the ban, the conversation shifted towards politics in sports in general, accompanied by a subtle shift in sentiment towards neutrality. Our evaluation serves as a framework for future research to explore the impact of sports activism and evolving public sentiment. This is especially useful in settings where labeling datasets for specific opinions is unfeasible, such as when events are unfolding.
How to Translate SQuAD to German? A Comparative Study of Answer Span Retrieval Methods for Question Answering Dataset Creation
Jens Kaiser, Agnieszka Falenska
This paper investigates the effectiveness of automatic span retrieval methods for translating SQuAD to German through a comparative analysis across two scenarios. First, we assume no gold-standard target data and find that TAR, a method using an alignment model, results in the highest question answering scores. Secondly, we switch to a scenario with a small target data and assess the impact of retrieval methods on fine-tuned models. Our results indicate that while fine-tuning generally enhances model performance, its effectiveness is dependent on the alignment of training and testing datasets.
LLM-based Translation Across 500 Years. The Case for Early New High German
Martin Volk, Dominic P. Fischer, Patricia Scheurer, Raphael Schwitter, Phillip Benjamin Ströbel
The recently developed large language models (LLMs) show surprising translation capabilities for modern languages. In contrast, this paper investigates the ability of GPT-4 and Gemini to translate 500-year-old letters from Early New High German into modern German. We experiment with a corpus from the 16th century that is partly in Latin and partly in ENH-German. This corpus consists of more than 3000 letters that have been edited and annotated by experts from the Institute of Swiss Reformation Studies. We exploit their annotations for the evaluation of machine translation from ENH-German to German. Our experiments show that using the lexical footnotes by the editors in the prompts or directly injected into the text leads to high quality translations.
Binary indexes for optimising corpus queries
Peter Ljunglöf, Nicholas Smallbone, Mijo Thoresson, Victor Salomonsson
To be able to search for patterns in annotated text corpora is crucial for many different research disciplines. However, searching for complex patterns in large corpora can take long time – sometimes several minutes or even hours.
We investigate how inverted indexes can be used for efficient searching in large annotated corpora, and in particular binary indexes. We show how corpus queries are translated into lookups in unary and binary inverted indexes, and give efficient strategies for combining the results using efficient set operations. In addition we discuss how to make use of binary indexes for more complex query types.
Leveraging Cross-Lingual Transfer Learning in Spoken Named Entity Recognition Systems
Moncef Benaicha, David Thulke, Mehmet Ali Tuğtekin Turan
Recent advancements in Named Entity Recognition (NER) have significantly enhanced text classification capabilities. This paper focuses on spoken NER, specifically aimed at spoken document retrieval, an area not widely studied due to the lack of comprehensive datasets for spoken contexts. Additionally, the potential for cross-lingual transfer learning in low-resource situations deserves further investigation. In our study, we applied transfer learning techniques across Dutch, English, and German using both pipeline and End-to-End (E2E) approaches. We employed Wav2Vec2 models on custom pseudo-annotated datasets to evaluate the adaptability of cross-lingual systems. Our exploration of different architectural configurations assessed the robustness of these systems in spoken NER. Results showed that the E2E model was superior to the pipeline model, particularly with limited annotation resources. Furthermore, transfer learning from German to Dutch improved performance by 7% over the standalone Dutch E2E system and 4% over the Dutch pipeline model. Our findings highlight the effectiveness of cross-lingual transfer in spoken NER and emphasize the need for additional data collection to improve these systems.
Evaluating and Fine-Tuning Retrieval-Augmented Language Models to Generate Text with Accurate Citations
Vinzent Penzkofer, Timo Baumann
Retrieval Augmented Generation (RAG) is becoming an essential tool for easily accessing large amounts of textual information. However, it is often challenging to determine whether the information in a given response originates from the context, the training, or is a result of hallucination. Our contribution in this area is twofold. Firstly, we demonstrate how existing datasets for information retrieval evaluation can be used to assess the ability of Large Language Models (LLMs) to correctly identify relevant sources. Our findings indicate that there are notable discrepancies in the performance of different current LLMs in this domain. Secondly, we utilise the datasets and metrics for citation evaluation to enhance the citation quality of small open-weight LLMs through fine-tuning. We achieve significant performance gains in this task, matching the results of much larger models.
Decoding 16th-Century Letters: From Topic Models to GPT-Based Keyword Mapping
Phillip Benjamin Ströbel, Stefan Aderhold, Ramona Roller
Probabilistic topic models for categorising or exploring large text corpora are notoriously difficult to interpret. Making sense of them has thus justifiably been compared to "reading tea leaves". Involving humans in labelling topics consisting of words is feasible but time-consuming, especially if one infers many topics from a text collection. Moreover, it is a cognitively demanding task, and domain knowledge might be required depending on the text corpus. We thus examine how using a Large Language Model (LLM) offers support in text classification. We compare how the LLM summarises topics produced by Latent Dirichlet Allocation, Non-negative Matrix Factorisation and BERTopic. We investigate which topic modelling technique provides the best representations by applying these models to a 16th-century correspondence corpus in Latin and Early New High German and inferring keywords from the topics in a low-resource setting. We experiment with including domain knowledge in the form of already existing keyword lists. Our main findings are that the LLM alone provides usable topis already. However, guiding the LLM towards what is expected benefits the interpretability. We further want to highlight that using nouns and proper nouns only makes for good topic representations.
Towards Improving ASR Outputs of Spontaneous Speech with LLMs
Karner Manuel, Julian Linke, Mark Kröll, Barbara Schuppler, Bernhard C Geiger
This paper presents ongoing work towards an initial understanding of how large language models (LLMs) can assist automatic speech recognition (ASR) tasks. More concretely, we investigate if LLMs can improve hypotheses obtained from ASR systems, and if so, which patterns in the hypothesis allow for a correction. Our results show that LLMs can mainly correct syntax errors or errors caused by ASR systems splitting long words. We further find that in the majority of cases the word error rates w.r.t. the human annotation increase when an LLM is applied, while at the same time the semantic similarity with the human annotation improves.
Exploring Data Acquisition Strategies for the Domain Adaptation of QA Models
Maurice Falk, Adrian Ulges, Dirk Krechel
Domain adaptation in Question-Answering (QA) is of importance when deploying models in new target domains where specific terminology and information needs exist. Adaptation commonly relies on a supervised fine-tuning using datasets composed of contexts, questions, and answers from the new domain. However, the annotation of such datasets is known to demand significant time and resources. In this work, a semi-automatic approach is investigated, where -- instead of a fully manual acquisition -- only answer spans (or questions, respectively) are selectively labeled, and a generative model provides a corresponding question (or answer). The efficacy of the proposed approach is compared against LLM-based auto-generative methods. Through experiments on diverse domain-specific QA datasets, both from the research community and industry practice, the superiority of the semi-automatic approach in obtaining higher QA performance is demonstrated.
Estimating Word Concreteness from Contextualized Embeddings
Christian Wartena
Concreteness is a property of words that has recently received attention in computational linguistics. Since concreteness is a property of word senses rather than of words, it makes most sense to determine concreteness in a given context. Recent approaches for predicting the concreteness of a word occurrence in context have relied on collecting many features from all words in the context. In this paper, we show that we can achieve state-of-the-art results by using only contextualized word embeddings of the target words. We circumvent the problem of missing training data for this task by training a regression model on context-independent concreteness judgments, which are widely available for English. The trained model needs only a few additional training data to give good results for predicting concreteness in context. We can even train the initial model on English data and do the final training on another language and obtain good results for that language as well.
Features and Detectability of German Texts Generated with Large Language Models
Verena Irrgang, Veronika Solopova, Steffen Zeiler, Robert M. Nickel, Dorothea Kolossa
The proliferation of generative language models poses significant challenges in distinguishing between human- and AI-generated texts. This study focuses on detecting German texts produced by various Large Language Models (LLMs). We investigated the impact of the training data composition on the model's ability to generalize across unknown genres and generators and still perform well on its test set.
Our study confirms that models trained on data from a single generator excel at detecting that very generator, but struggle to detect others. We expanded our analysis by considering correlations between linguistic features and results from explainable AI. The findings underscore that generator-specific approaches are likely necessary to enhance the accuracy and reliability of text generation detection systems in practical scenarios.
Our code can be found in the Github repository: https://github.com/vernsy/generated_text_detector
Exploring Automatic Text Simplification for Lithuanian
Justina Mandravickaitė, Egle Rimkiene, Danguolė Kalinauskaitė, Danguolė Kotryna Kapkan
The purpose of text simplification is to reduce the complexity of the text while retaining important information. This aspect is relevant for improving accessibility for a wide range of readers, e.g., those with cognitive disorders, non-native speakers, as well as children and the general public among others. We report experiments on text simplification for Lithuanian, focusing on simplifying texts of an administrative style to a plain language level to make it easier to understand for common people. We chose mT5 and mBART as foundational models and fine-tuned them for the text simplification task. Also, we tested ChatGPT for this task. We evaluated the outputs of these models quantitatively and qualitatively. All in all, mBART model appeared to be most effective for simplifying Lithuanian text, reaching the highest BLEU, ROUGE and BERTscore scores. Qualitative evaluation by assessing the simplicity, meaning retention and grammaticality of sentences simplified by our fine-tuned models, complemented the results of evaluation metrics' scores.
Large Language Models as Evaluators for Scientific Synthesis
Julia Evans, Jennifer D'Souza, Sören Auer
Our study explores how well the state-of-the-art Large Language Models (LLMs), like GPT-4 and Mistral, can assess the quality of scientific summaries or, more fittingly, scientific syntheses, comparing their evaluations to those of human annotators. We used a dataset of 100 research questions and their syntheses made by GPT-4 from abstracts of five related papers, checked against human quality ratings. The study evaluates both the closed-source GPT-4 and the open-source Mistral model's ability to rate these summaries and provide reasons for their judgments. Preliminary results show that LLMs can offer logical explanations that somewhat match the quality ratings, yet a deeper statistical analysis shows a weak correlation between LLM and human ratings, suggesting the potential and current limitations of LLMs in scientific synthesis evaluation.
Role-Playing LLMs in Professional Communication Training: The Case of Investigative Interviews with Children
Don Tuggener, Teresa Schneider, Ariana Huwiler, Tobias Kreienbühl, Simon Hischier, Pius von Däniken, Susanna Niehaus
We present a novel approach for professional communication training in which Large Language Models (LLMs) are guided to dynamically adapt to inappropriate communication techniques by producing false information that match the biased expectations of an interviewer. We achieve this by dynamically altering the LLM’s system prompt in conjunction with a classifier that detects undesirable communication behaviour. We develop this approach for training German speaking criminal investigators who interview children in alleged sexual abuse cases. We describe how our approach operationalises the strict communication requirements for such interviews and how it is integrated into a full, end-to-end learning environment that supports speech interaction with 3D virtual characters. We evaluate several aspects of this environment and report the positive results of an initial user study.
Analysing Effects of Inducing Gender Bias in Language Models
Stephanie Gross, Brigitte Krenn, Craig Lincoln, Lena Holzwarth
It is inevitable that language models are biased to a certain extent. There are two approaches to deal with bias: i) find mitigation strategies and ii) acquire knowledge about the existing bias in a language model, be explicit about it and its desired and undesired potential influence on a certain application. In this paper, we present an approach where we deliberately induce bias by pre-training an existing language model on different additional datasets, with the purpose of inducing a bias (gender bias) and a domain shift (social media, manosphere). We then use a novel, qualitative approach to show that gender bias (bias shift), and attitudes and stereotypes of the domain (domain shift) are also reflected in the words generated by the respective LM.
Exploring Phonetic Features in Language Embeddings for Unseen Language Varieties of Austrian German
Lorenz Gutscher, Michael Pucher
Vectorized language embeddings of raw audio data improve tasks like language recognition, automatic speech recognition, and machine translation. Although embeddings exhibit high effectiveness in their respective tasks, unraveling explicit information or meaning encapsulated within the embeddings proves challenging. This study investigates a multilingual model's ability to capture features from phonetic, articulatory, variety, and speaker categories from brief audio segments comprising five consecutive phones spoken by Austrian speakers. Within the employed model for extraction, German serves as one of the pre-trained languages used. However, the manner in which the model processes Austrian varieties presents an intriguing area for investigation. Using a k-nearest neighbor classifier, it is tested whether the encoded features are prominent in the embedding. While characteristics like variety are effectively classified, the accuracy of phone classification is particularly high for specific phones that are characteristic of the respective dialect/sociolect.
Word alignment in Discourse Representation Structure parsing
Christian Obereder, Gabor Recski
Discourse Representation Structures (DRS) are formal representations of linguistic semantics based on Discourse Representation Theory (DRT, Kamp et al. 2011) that represent meaning as conditions over discourse referents. State-of-the-art DRS parsers learn the task of mapping text to DRSs from annotated corpora such as the Parallel Meaning Bank (PMB, Abzianidze et al. 2017). Using DRS in downstream NLP applications such as Named Entity Recognition (NER), Relation Extraction (RE), or Open Information Extraction (OIE) requires that DRS clauses produced by a parser be aligned with words of the input sentence. We propose a set of methods for extending such models to learn DRS-to-word alignment in two ways, by using learned attention weights for alignment and by adding alignment information from the PMB to the training data. Our results demonstrate that combining the two methods can achieve an alignment accuracy of over 98%. We also perform manual error analysis, showing that most remaining alignment errors are caused by one-off mistakes, many of which occur in sentences with multi-word expressions.